MiniMax M2.7 Builds Self-Evolving Harness

Hot take: this feels like MiniMax trying to productize the model as part of the R&D loop, not just as an output engine. The biggest signal is the self-evolving framing: the model is being used to improve the harness that trains and evaluates it. The reported 30% to 50% share of RL team workflow is eye-catching, but it should be treated as a company claim until independent validation appears. MiniMax is clearly emphasizing agentic competence over generic chat quality, with claims around coding, tool search, memory, and multi-step task execution. The benchmark story is broad: software engineering, repo-level delivery, terminal/system understanding, and office tasks all get a mention. If the claims hold up, M2.7 looks like a credible step toward models that function as operational teammates inside engineering orgs.